In [29]:
from google.colab import drive
drive.mount('/content/gdrive')
Importing required libraries.¶
In [0]:
import numpy as np
import cv2
import imageio
import glob
import matplotlib.pyplot as plt
from tensorflow import keras
from tensorflow.keras import layers
from scipy import ndimage
Loading our dataset images.¶
In [0]:
path = r'/content/gdrive/My Drive/LOLdataset/our485/high'
all_files = glob.glob(path + "/*.png")
high = list()
all_files.sort()
for fileName in all_files:
img = imageio.imread(fileName)
high.append(img)
high = np.array(high)
path = r'/content/gdrive/My Drive/LOLdataset/our485/low'
all_files = glob.glob(path + "/*.png")
all_files.sort()
low = list()
for fileName in all_files:
img = imageio.imread(fileName)
low.append(img)
low = np.array(low)
viewing an example¶
In [6]:
plt.figure(figsize=(30,30))
plt.subplot(1,2,1)
plt.title("Ground Truth",fontsize=20)
plt.imshow(high[479])
plt.subplot(1,2,2)
plt.title("Low Light Image",fontsize=20)
plt.imshow(low[479])
Out[6]:
Deep Neural Network Model¶
We employ a plain CNN of seven convolutional layers with symmetrical concatenation. Each layer consists of 32 convolutional kernels of size 3x3 and stride 1 followed by the ReLU activation function. The last convolutional layer is followed by the sigmoid activation function which produce parameter map of pixel range [0:1]
In [7]:
inputs = keras.Input(shape=(None, None, 3), name='img')
out1 = layers.Conv2D(32, (3,3), activation='relu', padding="same", strides=(1, 1))(inputs)
out2 = layers.Conv2D(32, (3,3), activation='relu', padding="same", strides=(1, 1))(out1)
out3 = layers.Conv2D(32, (3,3), activation='relu', padding="same", strides=(1, 1))(out2)
out4 = layers.Conv2D(32, (3,3), activation='relu', padding="same", strides=(1, 1))(out3)
in5 = layers.add([out3, out4])
out5 = layers.Conv2D(32, (3,3), activation='relu', padding="same", strides=(1, 1))(in5)
in6 = layers.add([out2, out5])
out6 = layers.Conv2D(32, (3,3), activation='relu', padding="same", strides=(1, 1))(in6)
in7 = layers.add([out1, out6])
outputs = layers.Conv2D(1, (3,3), activation='sigmoid', padding="same", strides=(1, 1))(in7)
model = keras.Model(inputs, outputs)
model.compile(optimizer="adam", loss='mean_squared_error')
model.summary()
shape of layers.¶
In [8]:
keras.utils.plot_model(model, 'mini_resnet.png', show_shapes=True)
Out[8]:
Training the model¶
In [9]:
def GenerateInputs(X,y):
for i in range(len(X)):
X_input = X[i].reshape(1,400,600,3)
y_temp = y[i] / 255
y_input = y_temp.reshape(1,400,600,3)
yield (X_input,y_input)
model.fit(GenerateInputs(low,high), epochs=20, verbose=1, steps_per_epoch=24)
Out[9]:
Saving the trained model¶
In [0]:
model.save(r'/content/gdrive/My Drive/ImageProcessing/ourModel.h5')
Loading some low light images to test.¶
In [0]:
path = r'/content/gdrive/My Drive/LOLdataset/eval15/low'
all_files = glob.glob(path + "/*")
x = list()
all_files.sort()
for fileName in all_files:
img = imageio.imread(fileName)
x.append(img)
X = np.array(x)
we converted the steps of iterations into recursive function.
¶
inputs:
img: the low light image (I, Eold)
index: no of iterations
flag: default = 1
for more info. see README.md
In [0]:
def Enhance(img, index, flag):
if index == 0:
return img
elif flag == 1:
h, w, c = img.shape
test = model.predict(img.reshape(1, h, w, 3))
temp = img / 255
image = temp + ((test[0,:,:,:] * temp)*(1-temp))
index = index - 1
flag = 0
return Enhance(image, index, flag)
else:
h, w, c = img.shape
temp = model.predict(img.reshape(1, h, w, 3))
image = img + ((temp[0,:,:,:] * img)*(1-img))
index = index - 1
return Enhance(image, index, flag)
Examples¶
In [15]:
Image = X[1]
plt.figure(figsize=(30,30))
plt.subplot(1,2,1)
plt.title("Low Light Image",fontsize=20)
plt.imshow(Image)
plt.subplot(1,2,2)
plt.title("Enhanced Image",fontsize=20)
plt.imshow(Enhance(Image, 8, 1))
Out[15]:
In [17]:
Image = X[2]
plt.figure(figsize=(30,30))
plt.subplot(1,2,1)
plt.title("Low Light Image",fontsize=20)
plt.imshow(Image)
plt.subplot(1,2,2)
plt.title("Enhanced Image",fontsize=20)
plt.imshow(Enhance(Image, 8, 1))
Out[17]:
In [23]:
Image = X[6]
plt.figure(figsize=(30,30))
plt.subplot(1,2,1)
plt.title("Low Light Image",fontsize=20)
plt.imshow(Image)
plt.subplot(1,2,2)
plt.title("Enhanced Image",fontsize=20)
plt.imshow(Enhance(Image, 12, 1))
Out[23]:
Extreme level of low light images¶
In [0]:
path = r'/content/gdrive/My Drive/ImageProcessing/DataSet/test'
all_files = glob.glob(path + "/*")
x = list()
all_files.sort()
for fileName in all_files:
img = imageio.imread(fileName)
x.append(img)
Z = np.array(x)
Examples¶
In [27]:
Image = Z[0]
plt.figure(figsize=(30,30))
plt.subplot(1,2,1)
plt.title("Low Light Image",fontsize=20)
plt.imshow(Image)
plt.subplot(1,2,2)
plt.title("Enhanced Image",fontsize=20)
plt.imshow(Enhance(Image, 12, 1))
Out[27]:
In [28]:
Image = Z[3]
plt.figure(figsize=(30,30))
plt.subplot(1,2,1)
plt.title("Low Light Image",fontsize=20)
plt.imshow(Image)
plt.subplot(1,2,2)
plt.title("Enhanced Image",fontsize=20)
plt.imshow(Enhance(Image, 12, 1))
Out[28]:
